Supporting the Digital Humanities

Vienna, 19–20 October 2010

Findings and Outcomes of the musicSpace Project

Speaker: David Bretherton (D.Bretherton@soton.ac.uk)

Co-Authors: Daniel A. Smith, mc schraefel, Joe Lambert

Presentation overview

I am going to focus on one particular outcome of musicSpace: a successor project called ‘MusicNet’.

I will concentrate on how musicSpace provided the motivation for MusicNet

2

musicSpace

3

3-year project that concluded September 2010http://musicspace.mspace.fm

musicSpace’s goals

To integrate access to leading online music resources using the mSpace faceted browser.

Demonstrate that integration could support rapid exploration & knowledge building.

Enable complex, multipart queries.

4

MusicNet

5

July 2010 – June 2011http://musicnet.mspace.fm

MusicNet’s goals

Mint URIs for composers so that content providers can unambiguously identify them.

– Hope to expand to include all music-related entities.

Publish alignment data to back-link into our data partners’ catalogues, and to other resources.

Build a suite of tools to support the alignment and integration of new linked data resources.

Build a demonstration service to illustrate the uses and benefits of the URIs and alignment data.

6

Contents

1. Brief overview of musicSpace

2. How musicSpace provided the motivation for ‘MusicNet’

3. MusicNet’s alignment tool

7

1. Brief overview of musicSpace

8

Problem

9

10

Centuries of material ...

11

... is now increasingly digitised

Yet data is often ‘siloed’.

Geographical dispersal has been replaced by virtual dispersal on the web. Data is now segregated into countless online repositories by: – Media type (text, image, audio,

video)– Date of creation/publication– Subject

12

Yet data is often ‘siloed’.

Geographical dispersal has been replaced by virtual dispersal on the web. Data is now segregated into countless online repositories by: – Language– Copyright holder– Ad hoc/insecure nature of project

funding

13

Yet data is often ‘siloed’.

Interoperability has generally not been given a high enough priority.

14

Using current online music data resources presents barriers at

all stages of the research process:

15

It is hard to speculatively browse around a subject area.

‘Real-world’ multipart queries are effectively intractable.

16

The barriers to tractability and their solutions

Need to consult several sources … and metadata from one source cannot guide searches of another source.

Insufficient granularity of data and/or search option.

Multi-part queries have to be broken down and results collated manually.

Solutions:

Integration

Increase granularity

Optimally interactive UI (‘mSpace’)

Solution

17

18

‘musicSpace’ is a faceted browser

19

Demonstration

‘What recording of works by Cage exist, which performers have recorded a particular work by Cage, and what else by Cage have they recorded?

Screencast 1:

http://www.youtube.com/watch?v=keTN12OWies&hd=1

2. How musicSpace provided the motivation for MusicNet

20

Data is not ‘clean’...

21

Schubert Schubert, Franz Schubert, Franz Peter Shu-po-tʻe, ‡d  1797-1828 Schubert ‡d  1797-1828 F. P. Schubert Schubert, ... ‡d  1797-1828 Schubert, F. Schubert, F. ‡d  1797-1828 Schubert, Fr. Schubert, Fr. ‡d  1797-1828 Schubert, Franciszek. Schubert, Franc. ‡d  1797-1828 Schubert, Francois ‡d  1797-1828 Schubert, Franz P. ‡d  1797-1828

Schubert, Franz Peter Schubert, Franz Peter, ‡d  1797-1828 Schubert, Franz Peter ‡d  1797-1828 Schubert, Francois, ‡d  1797-1828 Schubert. Schubert ‡d  1797-1828 Shu-po-tʿe ‡d  1797-1828 Shubert, F. (Frant $s% ) ‡d  1797-1828 Shubert, F. ‡q  (Frant $s% ), ‡d  1797-1828 Shubert, Frant $s% , ‡d  1797-1828 Shubert, Frant $s% ‡d  1797-1828 Shūberuto, F. Shūberuto, Furantsu ‡d  1797-1828 Subert, Franc ‡d  1797-1828 Subertas, F. (Francas), ‡d  1797-1828

Subertas, Francas Peteris,   1797-1828‡d Subert, F.

, .Subertas F ‡d 1797-1828 פרנץ, שוברט

シューベルト, F., 1797-1828 シューベルト , フランツ ‡d  1797-1828 舒柏特 , 弗朗茨 Schubert, Francois   1797-1828‡d

, Schubert Franz Peter   1797-1828‡d

Causes of dirty data

Different naming conventions;– e.g. ‘Bach, Johann Sebastian’ or ‘J. S. Bach’

Inclusion of non-name data in name field; – e.g. ‘Schubert, Franz, 1797-1828. Songs’,

or ‘Allen, Betty (Teresa)’

Different languages (and alphabets);

User input errors. – e.g. ‘Bach, Johan Sebastien’

22

Dirty data degrades the user experience

23

Searching for compositions by the composer Franz Schubert (1797–1828)...

Screencast 2:

http://www.youtube.com/watch?v=pFsYfz1vlAg&hd=1

3. MusicNet’s alignment tool

24

Prototype 1 (musicSpace era)

25

Used Alignment API & Google Docs

We used Alignment API to compare the names as strings, using WordNet to enable word stemming, synonym support, etc.

Alignment API produces a similarity measure for each possible match.

We planned to set a threshold for automatic approval.

Matches below that threshold would be sent to a Google Docs spreadsheet for expert review.

26

Shortcoming 1: no threshold

It was not possible to identify a threshold for automatic approval.

Terms are judged to be similar if they have just, say, one different character, but a difference of one character is usually significant in a name.

Names are proper nouns, and so are unsuited to WordNet’s assumptions about misspelling.

27

Shortcoming 1: no threshold

False matches with high similarity measures:

True matches with low similarity measures:

28

Shortcoming 2: no context

Alignment API compares names as strings, and the system strips the names of their context (i.e. additional metadata). – Lack of context meant the musicologist had

no way to verify the match.

Significant flaw; automation had failed so we where relying on manual review.

29

Prototype 2 (building a custom tool

for MusicNet)

30

Lessons learned

From Prototype 1:– A completely automated solution is out of the

question (for the moment...). – We needed a custom tool with a human-friendly UI

(we also wanted keyboard shortcuts for speed).– Access to additional metadata (i.e. context), so

matches can be researched by the reviewer.

From experience with faceted browsers: – Alphabetically sorted columns enable one to spot

synonymous names at a glance.· Normally sources give names surname first; duplication

arises from the different representation of given names.

31

Alignment process Data*

32

Suggested groups

Algorithm compares hash of alpha-only l.c. version of name

No groups suggested

User verified* or rejected*

Synonym groups

Manual grouping (research*)

URIs Alternative names Back links*

UI of Prototype 2

33

Prototype 2 demo

34

Screencast 3:

http://www.youtube.com/watch?v=5f8iaryZMk0&hd=1

Indicative use cases

Composer URIs: – Music(ological) content providers– Basis of a (re)search portal

Alignment tool: – Aligning databases with no authorities;– Or where authorities are inconsistent.

35

36

Thank you for listening!